Hubness and Pollution: Delving into Cross-Space Mapping for Zero-Shot Learning

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 270–280, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Hubness and Pollution: Delving into Cross-Space Mapping for Zero-Shot Learning Angeliki Lazaridou Georgiana Dinu Marco Baroni Center for Mind/Brain Sciences University of Trento {angeliki.lazaridou|georgiana.dinu|marco.baroni}@unitn.it Abstract Zero-shot methods in language, vision and other domains rely on a cross-space map- ping function that projects vectors from the relevant feature space (e.g., visual- feature-based image representations) to a large semantic word space (induced in an unsupervised way from corpus data), where the entities of interest (e.g., objects images depict) are labeled with the words associated to the nearest neighbours of the mapped vectors. Zero-shot cross-space mapping methods hold great promise as a way to scale up annotation tasks well be- yond the labels in the training data (e.g., recognizing objects that were never seen in training). However, the current perfor- mance of cross-space mapping functions is still quite low, so that the strategy is not yet usable in practical applications. In this paper, we explore some general properties, both theoretical and empirical, of the cross-space mapping function, and we build on them to propose better meth- ods to estimate it. In this way, we attain large improvements over the state of the art, both in cross-linguistic (word trans- lation) and cross-modal (image labeling) zero-shot experiments. 1 Introduction In many supervised problems, the parameters of a classification function are estimated on (x, y) pairs, where x is a vector representing a training instance in some feature space, and y is the label assigned to the instance. For example, in image labeling x contains visual features extracted from a picture and y is the name of the object depicted in the picture (Grauman and Leibe, 2011). Since each label is treated as an unanalyzed primitive, this approach requires ad-hoc annotation for each label of interest, and it will not scale up to chal- lenges where the potential label set is vast (for ex- ample, bilingual dictionary induction, where the label set corresponds to the full vocabulary of the target language). Zero-shot methods (Palatucci et al., 2009) ad- dress the scalability problem by building on the observation that the labels of interest are often words (or longer linguistic expressions), which stand in a semantic similarity relation to each other. Moreover, distributional approaches allow us to estimate very large semantic word spaces in an efficient and unsupervised manner, using just unannotated text corpora as input (Turney and Pantel, 2010). Extensive evidence has shown that the similarity estimates obtained by representing words as vectors in such corpus-induced seman- tic spaces are extremely accurate (Baroni et al., 2014). Under the assumption that the domain of interest (e.g., objects in pictures, words in a source language) exhibits comparable similarity structure to that manifested in language, we can rephrase the learning task, from inducing multiple functions from the source feature space onto independent atomic labels, to that …

Content truncated to 500 words